🤖 MSME Scheme Assistant Chatbot

An AI-powered chatbot built using LLMs, sentence embeddings, and MongoDB to assist Indian MSMEs with discovering relevant government schemes. Also includes support for Telugu translations and PDF-based Q&A.

---

🧠 Features

• Udyam ID-based enterprise profiling
• Manual enterprise profile input
• LLM-generated search queries based on profile
• Embedding-based scheme recommendation
• Field-level answers (eligibility, benefits, documents, etc.) using LLM
• Telugu translation support via IndicTrans2
• Upload your own PDF and query it
• Built-in Gradio UI for seamless interaction

---

🛠️ Tech Stack

• 🤗 Transformers: LLMs & Tokenizers
• 🧠 SentenceTransformers: Semantic similarity
• 🗃️ MongoDB: Vector database for profiles, schemes, and uploaded PDFs
• 🧩 LangChain: Prompt templates
• 📚 IndicTrans2: English → Telugu translation
• 🖼️ Gradio: UI interface
• 🧮 PyMuPDF: PDF text extraction

---

🚀 Getting Started

🔧 Installation

pip install pymongo sentence-transformers torch transformers langchain_community pymupdf tools
pip install bitsandbytes scipy accelerate datasets sentencepiece

⚠️ Run in Google Colab for seamless PDF uploads using google.colab.files.

---

📁 MongoDB Setup

Ensure your MongoDB contains:

• udyam_profiles – MSME profile data
• schemes_chunks_only – Chunked scheme embeddings
• schemes_embedded – Full scheme info
• uploaded_pdf_temp – Temporary collection for uploaded PDFs

Update the mongo_uri in the script with your connection string.

---

🤖 LLMs Used

• Gemma-2B-IT base model + LoRA fine-tuned model: Vipplav/gemma-finetuned-faq
• Embedding model: BAAI/bge-small-en-v1.5
• Translator model: ai4bharat/indictrans2-en-indic-1B

---

🧪 Usage

🟢 Launch Chatbot

Run the script to start the Gradio interface.

💬 Chat Flow

1. Enter Udyam Registration Number or type manual
2. If manual:
   • Fill enterprise details interactively
3. Type show related schemes
4. Ask about eligibility, apply, or documents

🌐 Translate Response

• Click 🌐 Translate Last Scheme Reply for Telugu version

---

📄 PDF Chat

1. Upload a text-based PDF
2. Click 📄 Enable PDF Chat
3. Ask questions related to uploaded document
4. Use 🌐 Translate PDF Answer for Telugu translation

---

📦 File Structure

• final_msme_chatbot.py: End-to-end chatbot + PDF handler
• README.md: You're reading it

---

🧠 Logic Highlights

• Profile summarization → LLM prompt → Query
• Query embedding → MongoDB chunk similarity match
• Top schemes retrieved and displayed
• Further questions answered via LLM on stored chunk metadata
• Uploaded PDFs stored temporarily, chunked, and queried using top-k similarity

---

📝 Notes

• Only text-based PDFs are supported (no scanned images or heavy tables).
• Queries and translations are optimized for short, user-friendly outputs.
• You can switch between scheme-based chat and PDF chat seamlessly.

---

📍 TODOs / Improvements

• OCR for scanned PDFs (e.g., Tesseract)
• Streamlit-based deployment option
• Caching for repeated query embedding
• Feedback/rating collection

---

📜 License

MIT License